摘要
针对当前人脸签到效率偏低的问题,提出了一种基于单阶段的多人脸动态检测方案。首先输入包含人脸的图片,经过MobileNet卷积提取特征,接着送入FPN网络提取多尺度候选框,最后输入到上下文模块完成检测和分类。在实验对比过程中,将改进的单阶段网络与传统的目标检测网络以及卷积神经网络进行对比,改进后的方案因其出色的结构有效地减少了时间成本,在CPU环境中,在720P以及VGA视频总用时相比HOG方案快了至少2倍;在GPU环境中,对720P以及VGA视频相比CNN方案,检测速率分别提高了17.8%和1.3%。另外,在精度方面,检测的难度越大,提出的方案越有优势。当每帧人脸介于3~5个之间时,相较于HOG、Haar以及CNN方案,分别提高了约20.7%,18.4%和11.7%;当每帧人脸介于6~10个之间时,相较于HOG、Haar以及CNN方案分别提高了约52.5%,53.0%和34.4%。
In view of the low efficiency of face sign-in at present,a single-stage multiple-face dynamic detection scheme is proposed. The picture containing the human faces is inputted to extract the features of human faces by MobileNet convolution,and then sent into FPN(feature pyramid network) to extract the multi-scale candidate box,at last,inputted to the context module to complete the detection and classification. In the process of experimental comparison,the improved single-stage network was compared with the traditional object detection network and the convolutional neural network respectively. The improved scheme can effectively reduce the time cost because of its excellent structure. In the CPU environment,the total time consumption at 720 P and VGA video is at least twice as fast as that of the HOG(histogram of oriented gradient)scheme;in the GPU(graphics processing unit) environment,the detection rate for 720 P and VGA video is increased respectively by 17.8%and 1.3% in comparison with that of the CNN(convolutional neural network)scheme. In addition,in terms of the accuracy,the greater the difficulty of detection is,the more advantageous the proposed scheme shows. When the number of human faces in each frame is within 3~5,the detection accuracy of the proposed scheme is increased by about 20.7%,18.4% and 11.7%respectively in comparison with the HOG scheme,Haar scheme and CNN scheme;when the number of human faces in each frame is within 6~10, the detection accuracy of the proposed scheme is increased by about 52.5%, 53.0% and 34.4%respectively in comparison with HOG scheme,Haar scheme and CNN scheme.
作者
巩稼民
张凯泽
杨红蕊
赵梦凯
杨立春
GONG Jiamin;ZHANG Kaize;YANG Hongrui;ZHAO Mengkai;YANG Lichun(School of Communication and Information Engineering,Xi’an University of Posts and Telecommunications,Xi’an 710121,China;School of Electronic Engineering,Xi’an University of Posts and Telecommunications,Xi’an 710121,China;Shenzhen Penguin Network Technology Co.,Ltd.,Shenzhen 518000,China)
出处
《现代电子技术》
2021年第17期49-55,共7页
Modern Electronics Technique
基金
国家自然科学基金项目(61775180)
陕西省重点研发计划(2020KWZ-017)
校企共建研发项目:基于大数据的智慧教育自动监控与评估系统(204020204)。
